sglang 0.2.0__tar.gz → 0.2.2__tar.gz

{sglang-0.2.0/sglang.egg-info → sglang-0.2.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: sglang
-Version: 0.2.0
+Version: 0.2.2
 Summary: SGLang is yet another fast serving framework for large language models and vision language models.
 License:                                  Apache License
                                    Version 2.0, January 2004
@@ -249,7 +249,7 @@ Requires-Dist: sglang[litellm]; extra == "all"
 
 --------------------------------------------------------------------------------
 
-| [**Blog**](https://lmsys.org/blog/2024-01-17-sglang/) | [**Paper**](https://arxiv.org/abs/2312.07104) |
+| [**Blog**](https://lmsys.org/blog/2024-07-25-sglang-llama3/) | [**Paper**](https://arxiv.org/abs/2312.07104) |
 
 SGLang is a fast serving framework for large language models and vision language models.
 It makes your interaction with models faster and more controllable by co-designing the backend runtime and frontend language.
@@ -259,13 +259,14 @@ The core features include:
 - **Flexible Frontend Language**: Enables easy programming of LLM applications with chained generation calls, advanced prompting, control flow, multiple modalities, parallelism, and external interactions.
 
 ## News
-- [2024/04] 🔥 SGLang is used by the official **LLaVA-NeXT (video)** release ([blog](https://llava-vl.github.io/blog/2024-04-30-llava-next-video/)).
-- [2024/02] 🔥 SGLang enables **3x faster JSON decoding** with compressed finite state machine ([blog](https://lmsys.org/blog/2024-02-05-compressed-fsm/)).
-- [2024/01] SGLang provides up to **5x faster inference** with RadixAttention ([blog](https://lmsys.org/blog/2024-01-17-sglang/)).
+- [2024/07] 🔥 Faster Llama3 Serving with SGLang Runtime (vs. TensorRT-LLM, vLLM) ([blog](https://lmsys.org/blog/2024-07-25-sglang-llama3/)).
+- [2024/04] SGLang is used by the official **LLaVA-NeXT (video)** release ([blog](https://llava-vl.github.io/blog/2024-04-30-llava-next-video/)).
+- [2024/02] SGLang enables **3x faster JSON decoding** with compressed finite state machine ([blog](https://lmsys.org/blog/2024-02-05-compressed-fsm/)).
 
 <details>
 <summary>More</summary>
 
+- [2024/01] SGLang provides up to **5x faster inference** with RadixAttention ([blog](https://lmsys.org/blog/2024-01-17-sglang/)).
 - [2024/01] SGLang powers the serving of the official **LLaVA v1.6** release demo ([usage](https://github.com/haotian-liu/LLaVA?tab=readme-ov-file#demo)).
 
 </details>
@@ -302,7 +303,8 @@ pip install flashinfer -i https://flashinfer.ai/whl/cu121/torch2.3/
 ```
 
 ### Method 3: Using docker
-The docker images are available on Docker Hub as [lmsysorg/sglang](https://hub.docker.com/r/lmsysorg/sglang/tags).
+The docker images are available on Docker Hub as [lmsysorg/sglang](https://hub.docker.com/r/lmsysorg/sglang/tags), built from [Dockerfile](docker).
+Repalce `<secret>` below with your huggingface hub [token](https://huggingface.co/docs/hub/en/security-tokens).
 
 ```bash
 docker run --gpus all \
@@ -311,7 +313,7 @@ docker run --gpus all \
     --env "HUGGING_FACE_HUB_TOKEN=<secret>" \
     --ipc=host \
     lmsysorg/sglang:latest \
-    python3 -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B --host 0.0.0.0 --port 30000
+    python3 -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --host 0.0.0.0 --port 30000
 ```
 
 ### Common Notes
@@ -399,6 +401,21 @@ python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct
 - To enable fp8 quantization, you can add `--quantization fp8` on a fp16 checkpoint or directly load a fp8 checkpoint without specifying any arguments.
 - To enable experimental torch.compile support, you can add `--enable-torch-compile`. It accelerates small models on small batch sizes.
 
+### Run Llama 3.1 405B
+
+```bash
+# 2 nodes run 405B fp16
+# replace the `172.16.4.52:20000` with your own first node ip address and port, disable CUDA Graph temporarily
+# on the first node
+GLOO_SOCKET_IFNAME=eth0 python3 -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-405B-Instruct --tp 16 --nccl-init-addr 172.16.4.52:20000 --nnodes 2 --node-rank 0 --disable-cuda-graph --mem-frac 0.75
+
+# on the second
+GLOO_SOCKET_IFNAME=eth0 python3 -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-405B-Instruct --tp 16 --nccl-init-addr 172.16.4.52:20000 --nnodes 2 --node-rank 1 --disable-cuda-graph --mem-frac 0.75
+
+# single node run 405B fp8
+python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-405B-Instruct-FP8 --tp 8
+```
+
 ### Supported Models
 
 - Llama / Llama 2 / Llama 3 / Llama 3.1
@@ -656,15 +673,12 @@ for out in state.text_iter():
 - The `choices` argument in `sgl.gen` is implemented by computing the [token-length normalized log probabilities](https://blog.eleuther.ai/multiple-choice-normalization/) of all choices and selecting the one with the highest probability.
 - The `regex` argument in `sgl.gen` is implemented through autoregressive decoding with logit bias masking, according to the constraints set by the regex. It is compatible with `temperature=0` and `temperature != 0`.
 
-## Benchmark And Performance
-- Llama-7B on NVIDIA A10G, FP16, Tensor Parallelism=1
-![llama_7b](assets/llama_7b.jpg)
 
-- Mixtral-8x7B on NVIDIA A10G, FP16, Tensor Parallelism=8
-![mixtral_8x7b](assets/mixtral_8x7b.jpg)
+## Benchmark And Performance
+![8b_throughput](https://lmsys.org/images/blog/sglang_llama3/8b_throughput.svg)
+![70b_fp8_throughput](https://lmsys.org/images/blog/sglang_llama3/70b_fp8_throughput.svg)
 
-- Learn more about the above [results](docs/benchmark_results.md).
-- Synthetic latency and throughput benchmark [scripts](https://github.com/sgl-project/sglang/tree/main/benchmark/latency_throughput).
+Learn more at this [blog](https://lmsys.org/blog/2024-07-25-sglang-llama3/).
 
 ## Roadmap
 [Development Roadmap (2024 Q3)](https://github.com/sgl-project/sglang/issues/634)

{sglang-0.2.0 → sglang-0.2.2}/README.md

@@ -4,7 +4,7 @@
 
 --------------------------------------------------------------------------------
 
-| [**Blog**](https://lmsys.org/blog/2024-01-17-sglang/) | [**Paper**](https://arxiv.org/abs/2312.07104) |
+| [**Blog**](https://lmsys.org/blog/2024-07-25-sglang-llama3/) | [**Paper**](https://arxiv.org/abs/2312.07104) |
 
 SGLang is a fast serving framework for large language models and vision language models.
 It makes your interaction with models faster and more controllable by co-designing the backend runtime and frontend language.
@@ -14,13 +14,14 @@ The core features include:
 - **Flexible Frontend Language**: Enables easy programming of LLM applications with chained generation calls, advanced prompting, control flow, multiple modalities, parallelism, and external interactions.
 
 ## News
-- [2024/04] 🔥 SGLang is used by the official **LLaVA-NeXT (video)** release ([blog](https://llava-vl.github.io/blog/2024-04-30-llava-next-video/)).
-- [2024/02] 🔥 SGLang enables **3x faster JSON decoding** with compressed finite state machine ([blog](https://lmsys.org/blog/2024-02-05-compressed-fsm/)).
-- [2024/01] SGLang provides up to **5x faster inference** with RadixAttention ([blog](https://lmsys.org/blog/2024-01-17-sglang/)).
+- [2024/07] 🔥 Faster Llama3 Serving with SGLang Runtime (vs. TensorRT-LLM, vLLM) ([blog](https://lmsys.org/blog/2024-07-25-sglang-llama3/)).
+- [2024/04] SGLang is used by the official **LLaVA-NeXT (video)** release ([blog](https://llava-vl.github.io/blog/2024-04-30-llava-next-video/)).
+- [2024/02] SGLang enables **3x faster JSON decoding** with compressed finite state machine ([blog](https://lmsys.org/blog/2024-02-05-compressed-fsm/)).
 
 <details>
 <summary>More</summary>
 
+- [2024/01] SGLang provides up to **5x faster inference** with RadixAttention ([blog](https://lmsys.org/blog/2024-01-17-sglang/)).
 - [2024/01] SGLang powers the serving of the official **LLaVA v1.6** release demo ([usage](https://github.com/haotian-liu/LLaVA?tab=readme-ov-file#demo)).
 
 </details>
@@ -57,7 +58,8 @@ pip install flashinfer -i https://flashinfer.ai/whl/cu121/torch2.3/
 ```
 
 ### Method 3: Using docker
-The docker images are available on Docker Hub as [lmsysorg/sglang](https://hub.docker.com/r/lmsysorg/sglang/tags).
+The docker images are available on Docker Hub as [lmsysorg/sglang](https://hub.docker.com/r/lmsysorg/sglang/tags), built from [Dockerfile](docker).
+Repalce `<secret>` below with your huggingface hub [token](https://huggingface.co/docs/hub/en/security-tokens).
 
 ```bash
 docker run --gpus all \
@@ -66,7 +68,7 @@ docker run --gpus all \
     --env "HUGGING_FACE_HUB_TOKEN=<secret>" \
     --ipc=host \
     lmsysorg/sglang:latest \
-    python3 -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B --host 0.0.0.0 --port 30000
+    python3 -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --host 0.0.0.0 --port 30000
 ```
 
 ### Common Notes
@@ -154,6 +156,21 @@ python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct
 - To enable fp8 quantization, you can add `--quantization fp8` on a fp16 checkpoint or directly load a fp8 checkpoint without specifying any arguments.
 - To enable experimental torch.compile support, you can add `--enable-torch-compile`. It accelerates small models on small batch sizes.
 
+### Run Llama 3.1 405B
+
+```bash
+# 2 nodes run 405B fp16
+# replace the `172.16.4.52:20000` with your own first node ip address and port, disable CUDA Graph temporarily
+# on the first node
+GLOO_SOCKET_IFNAME=eth0 python3 -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-405B-Instruct --tp 16 --nccl-init-addr 172.16.4.52:20000 --nnodes 2 --node-rank 0 --disable-cuda-graph --mem-frac 0.75
+
+# on the second
+GLOO_SOCKET_IFNAME=eth0 python3 -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-405B-Instruct --tp 16 --nccl-init-addr 172.16.4.52:20000 --nnodes 2 --node-rank 1 --disable-cuda-graph --mem-frac 0.75
+
+# single node run 405B fp8
+python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-405B-Instruct-FP8 --tp 8
+```
+
 ### Supported Models
 
 - Llama / Llama 2 / Llama 3 / Llama 3.1
@@ -411,15 +428,12 @@ for out in state.text_iter():
 - The `choices` argument in `sgl.gen` is implemented by computing the [token-length normalized log probabilities](https://blog.eleuther.ai/multiple-choice-normalization/) of all choices and selecting the one with the highest probability.
 - The `regex` argument in `sgl.gen` is implemented through autoregressive decoding with logit bias masking, according to the constraints set by the regex. It is compatible with `temperature=0` and `temperature != 0`.
 
-## Benchmark And Performance
-- Llama-7B on NVIDIA A10G, FP16, Tensor Parallelism=1
-![llama_7b](assets/llama_7b.jpg)
 
-- Mixtral-8x7B on NVIDIA A10G, FP16, Tensor Parallelism=8
-![mixtral_8x7b](assets/mixtral_8x7b.jpg)
+## Benchmark And Performance
+![8b_throughput](https://lmsys.org/images/blog/sglang_llama3/8b_throughput.svg)
+![70b_fp8_throughput](https://lmsys.org/images/blog/sglang_llama3/70b_fp8_throughput.svg)
 
-- Learn more about the above [results](docs/benchmark_results.md).
-- Synthetic latency and throughput benchmark [scripts](https://github.com/sgl-project/sglang/tree/main/benchmark/latency_throughput).
+Learn more at this [blog](https://lmsys.org/blog/2024-07-25-sglang-llama3/).
 
 ## Roadmap
 [Development Roadmap (2024 Q3)](https://github.com/sgl-project/sglang/issues/634)

{sglang-0.2.0 → sglang-0.2.2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "sglang"
-version = "0.2.0"
+version = "0.2.2"
 description = "SGLang is yet another fast serving framework for large language models and vision language models."
 readme = "README.md"
 requires-python = ">=3.8"

{sglang-0.2.0 → sglang-0.2.2}/sglang/bench_serving.py

@@ -369,7 +369,7 @@ def sample_random_requests(
 ) -> List[Tuple[str, int, int]]:
 
     input_lens = np.random.randint(
-        int(input_len * range_ratio),
+        max(int(input_len * range_ratio), 1),
         input_len + 1,
         size=num_prompts,
     )
@@ -415,7 +415,7 @@ def sample_random_requests(
             prompt_token_ids = tokenizer(prompt).input_ids
             prompt_len = len(prompt_token_ids)
 
-            if prompt_len <= input_lens[i]:
+            if prompt_len > input_lens[i]:
                 input_ids = prompt_token_ids[: input_lens[i]]
             else:
                 ratio = (input_lens[i] + prompt_len - 1) // prompt_len
@@ -935,7 +935,7 @@ if __name__ == "__main__":
     parser.add_argument(
         "--random-range-ratio",
         type=float,
-        default=1.0,
+        default=0.0,
         help="Range of sampled ratio of input/output length, "
         "used only for random dataset.",
     )

{sglang-0.2.0 → sglang-0.2.2}/sglang/global_config.py

@@ -17,7 +17,7 @@ class GlobalConfig:
 
         # Runtime constants: New generation token ratio estimation
         self.init_new_token_ratio = 0.7
-        self.base_min_new_token_ratio = 0.2
+        self.base_min_new_token_ratio = 0.1
         self.new_token_ratio_decay = 0.001
         self.new_token_ratio_recovery = 0.05
 

{sglang-0.2.0 → sglang-0.2.2}/sglang/srt/managers/controller/model_runner.py

@@ -121,7 +121,7 @@ class ModelRunner:
             skip_tokenizer_init=True,
         )
 
-        if is_llama3_405b_fp8(self.model_config):
+        if is_llama3_405b_fp8(self.model_config) and self.tp_size <= 8:
             # A temporary hack to fix the num_heads for meta-llama/Meta-Llama-3.1-405B-FP8 checkpoints
             self.model_config.hf_config.num_key_value_heads = 8
             vllm_model_config.hf_config.num_key_value_heads = 8

{sglang-0.2.0 → sglang-0.2.2}/sglang/srt/managers/io_struct.py

@@ -40,7 +40,10 @@ class GenerateReqInput:
             self.text is not None and self.input_ids is not None
         ):
             raise ValueError("Either text or input_ids should be provided.")
-        if self.sampling_params.get("n", 1) != 1:
+        if (
+            isinstance(self.sampling_params, dict)
+            and self.sampling_params.get("n", 1) != 1
+        ):
             is_single = False
         else:
             if self.text is not None:

{sglang-0.2.0 → sglang-0.2.2}/sglang/srt/openai_api/adapter.py

@@ -94,9 +94,14 @@ def load_chat_template_for_openai_api(chat_template_arg):
 async def v1_completions(tokenizer_manager, raw_request: Request):
     request_json = await raw_request.json()
     request = CompletionRequest(**request_json)
+    prompt = request.prompt
+    if isinstance(prompt, str) or isinstance(prompt[0], str):
+        prompt_kwargs = {"text": prompt}
+    else:
+        prompt_kwargs = {"input_ids": prompt}
 
     adapted_request = GenerateReqInput(
-        text=request.prompt,
+        **prompt_kwargs,
         sampling_params={
             "temperature": request.temperature,
             "max_new_tokens": request.max_tokens,

{sglang-0.2.0 → sglang-0.2.2}/sglang/srt/utils.py

@@ -626,6 +626,7 @@ def is_llama3_405b_fp8(model_config):
         and model_config.hf_config.intermediate_size == 53248
         and model_config.hf_config.num_hidden_layers == 126
         and model_config.hf_config.num_key_value_heads == 16
+        and hasattr(model_config.hf_config, "quantization_config")
         and model_config.hf_config.quantization_config["quant_method"] == "fbgemm_fp8"
     ):
         return True

sglang-0.2.2/sglang/version.py

@@ -0,0 +1 @@
+__version__ = "0.2.2"

{sglang-0.2.0 → sglang-0.2.2/sglang.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: sglang
-Version: 0.2.0
+Version: 0.2.2
 Summary: SGLang is yet another fast serving framework for large language models and vision language models.
 License:                                  Apache License
                                    Version 2.0, January 2004
@@ -249,7 +249,7 @@ Requires-Dist: sglang[litellm]; extra == "all"
 
 --------------------------------------------------------------------------------
 
-| [**Blog**](https://lmsys.org/blog/2024-01-17-sglang/) | [**Paper**](https://arxiv.org/abs/2312.07104) |
+| [**Blog**](https://lmsys.org/blog/2024-07-25-sglang-llama3/) | [**Paper**](https://arxiv.org/abs/2312.07104) |
 
 SGLang is a fast serving framework for large language models and vision language models.
 It makes your interaction with models faster and more controllable by co-designing the backend runtime and frontend language.
@@ -259,13 +259,14 @@ The core features include:
 - **Flexible Frontend Language**: Enables easy programming of LLM applications with chained generation calls, advanced prompting, control flow, multiple modalities, parallelism, and external interactions.
 
 ## News
-- [2024/04] 🔥 SGLang is used by the official **LLaVA-NeXT (video)** release ([blog](https://llava-vl.github.io/blog/2024-04-30-llava-next-video/)).
-- [2024/02] 🔥 SGLang enables **3x faster JSON decoding** with compressed finite state machine ([blog](https://lmsys.org/blog/2024-02-05-compressed-fsm/)).
-- [2024/01] SGLang provides up to **5x faster inference** with RadixAttention ([blog](https://lmsys.org/blog/2024-01-17-sglang/)).
+- [2024/07] 🔥 Faster Llama3 Serving with SGLang Runtime (vs. TensorRT-LLM, vLLM) ([blog](https://lmsys.org/blog/2024-07-25-sglang-llama3/)).
+- [2024/04] SGLang is used by the official **LLaVA-NeXT (video)** release ([blog](https://llava-vl.github.io/blog/2024-04-30-llava-next-video/)).
+- [2024/02] SGLang enables **3x faster JSON decoding** with compressed finite state machine ([blog](https://lmsys.org/blog/2024-02-05-compressed-fsm/)).
 
 <details>
 <summary>More</summary>
 
+- [2024/01] SGLang provides up to **5x faster inference** with RadixAttention ([blog](https://lmsys.org/blog/2024-01-17-sglang/)).
 - [2024/01] SGLang powers the serving of the official **LLaVA v1.6** release demo ([usage](https://github.com/haotian-liu/LLaVA?tab=readme-ov-file#demo)).
 
 </details>
@@ -302,7 +303,8 @@ pip install flashinfer -i https://flashinfer.ai/whl/cu121/torch2.3/
 ```
 
 ### Method 3: Using docker
-The docker images are available on Docker Hub as [lmsysorg/sglang](https://hub.docker.com/r/lmsysorg/sglang/tags).
+The docker images are available on Docker Hub as [lmsysorg/sglang](https://hub.docker.com/r/lmsysorg/sglang/tags), built from [Dockerfile](docker).
+Repalce `<secret>` below with your huggingface hub [token](https://huggingface.co/docs/hub/en/security-tokens).
 
 ```bash
 docker run --gpus all \
@@ -311,7 +313,7 @@ docker run --gpus all \
     --env "HUGGING_FACE_HUB_TOKEN=<secret>" \
     --ipc=host \
     lmsysorg/sglang:latest \
-    python3 -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B --host 0.0.0.0 --port 30000
+    python3 -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --host 0.0.0.0 --port 30000
 ```
 
 ### Common Notes
@@ -399,6 +401,21 @@ python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct
 - To enable fp8 quantization, you can add `--quantization fp8` on a fp16 checkpoint or directly load a fp8 checkpoint without specifying any arguments.
 - To enable experimental torch.compile support, you can add `--enable-torch-compile`. It accelerates small models on small batch sizes.
 
+### Run Llama 3.1 405B
+
+```bash
+# 2 nodes run 405B fp16
+# replace the `172.16.4.52:20000` with your own first node ip address and port, disable CUDA Graph temporarily
+# on the first node
+GLOO_SOCKET_IFNAME=eth0 python3 -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-405B-Instruct --tp 16 --nccl-init-addr 172.16.4.52:20000 --nnodes 2 --node-rank 0 --disable-cuda-graph --mem-frac 0.75
+
+# on the second
+GLOO_SOCKET_IFNAME=eth0 python3 -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-405B-Instruct --tp 16 --nccl-init-addr 172.16.4.52:20000 --nnodes 2 --node-rank 1 --disable-cuda-graph --mem-frac 0.75
+
+# single node run 405B fp8
+python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-405B-Instruct-FP8 --tp 8
+```
+
 ### Supported Models
 
 - Llama / Llama 2 / Llama 3 / Llama 3.1
@@ -656,15 +673,12 @@ for out in state.text_iter():
 - The `choices` argument in `sgl.gen` is implemented by computing the [token-length normalized log probabilities](https://blog.eleuther.ai/multiple-choice-normalization/) of all choices and selecting the one with the highest probability.
 - The `regex` argument in `sgl.gen` is implemented through autoregressive decoding with logit bias masking, according to the constraints set by the regex. It is compatible with `temperature=0` and `temperature != 0`.
 
-## Benchmark And Performance
-- Llama-7B on NVIDIA A10G, FP16, Tensor Parallelism=1
-![llama_7b](assets/llama_7b.jpg)
 
-- Mixtral-8x7B on NVIDIA A10G, FP16, Tensor Parallelism=8
-![mixtral_8x7b](assets/mixtral_8x7b.jpg)
+## Benchmark And Performance
+![8b_throughput](https://lmsys.org/images/blog/sglang_llama3/8b_throughput.svg)
+![70b_fp8_throughput](https://lmsys.org/images/blog/sglang_llama3/70b_fp8_throughput.svg)
 
-- Learn more about the above [results](docs/benchmark_results.md).
-- Synthetic latency and throughput benchmark [scripts](https://github.com/sgl-project/sglang/tree/main/benchmark/latency_throughput).
+Learn more at this [blog](https://lmsys.org/blog/2024-07-25-sglang-llama3/).
 
 ## Roadmap
 [Development Roadmap (2024 Q3)](https://github.com/sgl-project/sglang/issues/634)

sglang-0.2.0/sglang/version.py

@@ -1 +0,0 @@
-__version__ = "0.2.0"